A method, device, electronic device and storage medium for predicting click behavior

Abstract

The present disclosure relates to a click behavior prediction method, device, electronic device and storage medium, the method comprising: obtaining a first-order feature vector and a second-order cross feature vector corresponding to the feature information of an account and the feature information of a target multimedia resource; inputting the second-order cross feature vector into a click behavior prediction network of a click behavior prediction model, and outputting an initial prediction vector; inputting the first-order feature vector into a feature weight layer, and obtaining a target weight matrix corresponding to the first-order feature vector, and outputting a weighted first-order feature scalar according to the target weight matrix; splicing the initial prediction vector and the weighted first-order feature scalar and inputting them into the output layer of the click behavior prediction model to obtain a click behavior prediction result. The weight parameters in this method can adapt to different data distributions, and improve the model's ability to remember feature data without increasing the computational complexity of the model too much, thereby further improving the accuracy of the prediction.

Description

Technical Field

[0001] The present disclosure relates to the field of the Internet, and in particular to a click behavior prediction method, device, electronic device, and storage medium. Background Art

[0002] Depth-based click behavior prediction models have been widely used in the field of click behavior prediction in various industries and platforms.

[0003] During their research into related technologies, the inventors discovered that existing click behavior prediction models face two challenges. First, many deployments and algorithms designed to improve model performance come at the expense of increasing the model's structural or algorithmic complexity, which increases the computational effort and computational time required for model predictions. Second, due to the relatively small amount of offline training data and the large amount of online training data, coupled with the influence of time factors such as holidays and festivals, the data distributions for offline and online training vary. This results in the offline-trained model being unable to adapt well to the varying data distributions when predicting online data, resulting in poor prediction accuracy. Summary of the Invention

[0004] The present disclosure provides a click behavior prediction method, device, electronic device, and storage medium to at least address the problem in related technologies of being unable to fully capture crash information and increasing system load. The technical solution of the present disclosure is as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, a click behavior prediction method is provided, comprising:

[0006] Acquiring characteristic information of an account and characteristic information of a target multimedia resource to be recommended to the account;

[0007] Invoking a click behavior prediction model, the click behavior prediction model comprising at least a click behavior prediction network and a feature weight layer; the feature weight layer comprising a weight matrix with adjustable weight parameters;

[0008] Constructing a first-order feature vector and a second-order cross feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource;

[0009] Inputting the second-order cross feature vector into the click behavior prediction network and outputting an initial prediction vector;

[0010] Inputting the first-order eigenvector into the feature weight layer, obtaining a target weight matrix corresponding to the first-order eigenvector, and outputting a weighted first-order eigenscalar according to the target weight matrix;

[0011] Outputting a predicted click rate of the account for the target multimedia resource according to the initial prediction vector and the weighted first-order feature scalar.

[0012] Optionally, the click behavior prediction model is trained by the following method:

[0013] Obtaining an initial model; the initial model includes an initial click behavior prediction network and an initial feature weight layer;

[0014] Acquire multiple training samples, wherein the training samples include feature data of sample accounts and feature data of sample multimedia resources, and click data of the sample accounts on the sample multimedia resources;

[0015] Using the second-order cross feature vector of the training sample and the click data, the initial click behavior prediction network is trained to obtain the click behavior prediction network;

[0016] Training the initial feature weight layer using the first-order feature vector of the training sample, the click behavior prediction network, and the click data to obtain the feature weight layer;

[0017] A click behavior prediction model is obtained based on the click behavior prediction network and the feature weight layer.

[0018] Optionally, the initial feature weight layer includes an initial weight matrix, and the training of the initial feature weight layer using the training samples and the click behavior prediction network includes:

[0019] Obtaining first-order feature vectors of feature information of the sample account and feature information of the sample multimedia resource;

[0020] Determine a one-dimensional scalar corresponding to the first-order eigenvector, and determine a normalized scalar corresponding to the one-dimensional scalar;

[0021] The product of the normalized scalar and the preset initial weight matrix is used as the initial weighted first-order characteristic scalar;

[0022] Inputting the initial weighted first-order feature scalar and the output result of the click behavior prediction network into the output layer of the initial model to obtain an initial click behavior prediction result;

[0023] Based on the click data and the initial click behavior prediction result, the initial weight matrix is dynamically adjusted to obtain the feature weight layer.

[0024] Optionally, determining the one-dimensional scalar corresponding to the first-order eigenvector and determining a normalized scalar corresponding to the one-dimensional scalar includes:

[0025] Performing a sum operation on the first-order eigenvector to obtain a one-dimensional scalar corresponding to the first-order eigenvector;

[0026] Normalization is performed on the one-dimensional scalar to obtain a normalized scalar corresponding to the one-dimensional scalar; the normalized scalar is a scalar with a mean of 0 and a variance of 1.

[0027] Optionally, the click behavior prediction model further includes an embedding layer and a perception layer, and constructing a first-order feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource includes:

[0028] dividing the feature information of the account and the feature information of the target multimedia resource into discrete features and continuous features;

[0029] Inputting the discrete features into the embedding layer to obtain a first dense vector;

[0030] splicing the continuous features together and inputting them into the perception layer to obtain a second dense vector with the same dimension as the first dense vector;

[0031] The first dense vector and the second dense vector are concatenated to obtain a first-order feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource.

[0032] Optionally, the click behavior prediction model further includes an inner product layer, and constructing a second-order cross feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource includes:

[0033] The first-order feature vector is input into the inner product layer, and the inner product layer is used to calculate the inner product of every two first-order feature vectors to obtain a second-order cross feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource.

[0034] According to a second aspect of an embodiment of the present disclosure, a click behavior prediction device is provided, the device comprising:

[0035] a feature information acquisition module configured to acquire feature information of an account and feature information of a target multimedia resource to be recommended to the account;

[0036] A calling module is configured to execute a call to a click behavior prediction model, wherein the click behavior prediction model includes at least a click behavior prediction network and a feature weight layer; the feature weight layer includes a weight matrix with adjustable weight parameters;

[0037] A construction module is configured to construct a first-order feature vector and a second-order cross feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource;

[0038] An intermediate prediction module is configured to input the second-order cross feature vector into the click behavior prediction network and output an initial prediction vector;

[0039] a weighting module configured to input the first-order eigenvector into the feature weight layer, obtain a target weight matrix corresponding to the first-order eigenvector, and output a weighted first-order feature scalar according to the target weight matrix;

[0040] The click behavior prediction module is configured to output the predicted click rate of the account for the target multimedia resource based on the initial prediction vector and the weighted first-order feature scalar.

[0041] Optionally, the apparatus further includes a click behavior prediction model training module, wherein the training module is configured to execute:

[0042] Obtaining an initial model; the initial model includes an initial click behavior prediction network and an initial feature weight layer;

[0043] Acquire multiple training samples, wherein the training samples include feature data of sample accounts and feature data of sample multimedia resources, and click data of the sample accounts on the sample multimedia resources;

[0044] Using the second-order cross feature vector of the training sample and the click data, the initial click behavior prediction network is trained to obtain the click behavior prediction network;

[0045] Training the initial feature weight layer using the first-order feature vector of the training sample, the click behavior prediction network, and the click data to obtain the feature weight layer;

[0046] A click behavior prediction model is obtained based on the click behavior prediction network and the feature weight layer.

[0047] Optionally, the initial feature weight layer includes an initial weight matrix, and the training module is specifically configured to perform:

[0048] Obtaining first-order feature vectors of feature information of the sample account and feature information of the sample multimedia resource;

[0049] Determine a one-dimensional scalar corresponding to the first-order eigenvector, and determine a normalized scalar corresponding to the one-dimensional scalar;

[0050] The product of the normalized scalar and the preset initial weight matrix is used as the initial weighted first-order characteristic scalar;

[0051] Inputting the initial weighted first-order feature scalar and the output result of the click behavior prediction network into the output layer of the initial model to obtain an initial click behavior prediction result;

[0052] Based on the click data and the initial click behavior prediction result, the initial weight matrix is dynamically adjusted to obtain the feature weight layer.

[0053] Optionally, the training module is specifically configured to execute:

[0054] Performing a sum operation on the first-order eigenvector to obtain a one-dimensional scalar corresponding to the first-order eigenvector;

[0055] Normalization is performed on the one-dimensional scalar to obtain a normalized scalar corresponding to the one-dimensional scalar; the normalized scalar is a scalar with a mean of 0 and a variance of 1.

[0056] Optionally, the click behavior prediction model further includes an embedding layer and a perception layer, and the building module is specifically configured to perform:

[0057] dividing the feature information of the account and the feature information of the target multimedia resource into discrete features and continuous features;

[0058] Inputting the discrete features into the embedding layer to obtain a first dense vector;

[0059] splicing the continuous features together and inputting them into the perception layer to obtain a second dense vector with the same dimension as the first dense vector;

[0060] The first dense vector and the second dense vector are concatenated to obtain a first-order feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource.

[0061] Optionally, the click behavior prediction model further includes an inner product layer, and the construction module is specifically configured to perform:

[0062] The first-order feature vector is input into the inner product layer, and the inner product layer is used to calculate the inner product of every two first-order feature vectors to obtain a second-order cross feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource.

[0063] According to a third aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0064] processor;

[0065] a memory for storing instructions executable by the processor;

[0066] The processor is configured to execute the instructions to implement the click behavior prediction method as described in the first aspect.

[0067] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When instructions in the computer-readable storage medium are executed by a processor of a server, the server is enabled to execute the click behavior prediction method as described in the first aspect.

[0068] According to a fifth aspect of an embodiment of the present disclosure, a computer program product is provided, comprising a computer program / instruction, wherein the computer program / instruction, when executed by a processor, implements the click behavior prediction method described in the first aspect.

[0069] The technical solutions provided by the embodiments of the present disclosure bring at least the following beneficial effects:

[0070] This solution inputs the second-order cross feature vector into the click behavior prediction network, which reduces the input data dimension and the amount of calculation. In addition, this solution adds a feature weight layer and inputs the first-order feature vector into the feature weight layer. On the one hand, the first-order feature vector does not undergo complex cross-linking and can directly remember the characteristics of the data, thereby improving the model's ability to remember feature data. On the other hand, the feature weight layer can adjust the weight parameters according to the changes in the first-order feature vector, so that the model can adapt well to different data distributions. The increased computational complexity and spatial complexity of this model are both linear. Therefore, this solution improves the model's ability to remember feature data without increasing the model's computational complexity too much, thereby further improving the accuracy of the prediction.

[0071] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The accompanying drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute an improper limitation of the present disclosure.

[0073] Figure 1 is a flowchart showing a method for predicting click behavior according to an exemplary embodiment;

[0074] Figure 2 is a schematic diagram of a click behavior prediction method provided according to an exemplary embodiment;

[0075] Figure 3 is a flowchart of another click behavior prediction method provided according to an exemplary embodiment;

[0076] Figure 4is a flowchart of another click behavior prediction method provided according to an exemplary embodiment;

[0077] Figure 5 is a flowchart of a method for training a click behavior prediction model according to an exemplary embodiment;

[0078] Figure 6 is a flowchart of another method for training a click behavior prediction model according to an exemplary embodiment;

[0079] Figure 7 is a structural block diagram of a click behavior prediction device according to an exemplary embodiment;

[0080] Figure 8 The figure is a block diagram of an electronic device for click behavior prediction according to an exemplary embodiment. DETAILED DESCRIPTION

[0081] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0082] It should be noted that the terms "first," "second," and the like in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the numbers used in this manner are interchangeable where appropriate so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of apparatus and methods consistent with certain aspects of the present disclosure as detailed in the appended claims.

[0083] Figure 1 The figure is a flowchart showing the steps of a click behavior prediction method according to an exemplary embodiment.

[0084] In step S11 , a first-order feature vector and a second-order cross feature vector corresponding to feature information of an account and feature information of a target multimedia resource are obtained; the target multimedia resource is a multimedia resource to be recommended to the account.

[0085] Multimedia resources can be at least one of short videos, live broadcasts, advertisements, and webpage merchandise. When a server receives a request from an account to obtain multimedia resources, it triggers the server to obtain multiple multimedia resources to be recommended to the account from the current resources, thereby obtaining multiple target multimedia resources. In order to determine the multimedia resources ultimately distributed to the account from the target multimedia resources, the click behavior prediction method provided in an embodiment of the present invention is used to predict the account's click-through rate for each target multimedia resource. When the predicted click-through rate is higher than a corresponding threshold, the corresponding target multimedia resource can be exposed to the account.

[0086] The characteristic information of an account may include the account ID, age, gender, city, favorite content tags, disliked content tags, average viewing time, device ID, etc. The characteristic information of a target multimedia resource may include the target multimedia resource ID, resource category, content tags, duration, etc.

[0087] The characteristic information of the account and the characteristic information of the target multimedia resource are encoded and mapped into a vector, which is a first-order characteristic vector. All characteristic information of an account corresponds to a first-order characteristic vector, and all characteristic information of a target multimedia resource also corresponds to a first-order characteristic vector.

[0088] Each pair of first-order eigenvectors is combined to obtain multiple second-order cross eigenvectors. The combination method may include multiplication or inner product.

[0089] In step S12, the second-order cross feature vector is input into a click behavior prediction network of a click behavior prediction model, and an initial prediction vector of the account's click behavior on the target multimedia resource is output.

[0090] The click behavior prediction model is a pre-trained model. The server predicts the account's click behavior on the target multimedia resource by calling the click behavior prediction network in the click behavior prediction model.

[0091] Specifically, the second-order cross-feature vector obtained in step S11 is used as input to the click behavior prediction network to obtain the prediction result of the click behavior prediction network. This prediction result is in vector form and is not the final prediction result. It is used as the initial prediction vector and combined with the weighted first-order feature scalar in the following steps to obtain the final prediction result.

[0092] In one embodiment, the click behavior prediction network is a DNN (Deep Neural Networks), and the prediction method is shown in formula (1):

[0093]

[0094] Among them, x dnn is the output of DNN, is the learnable parameter of DNN, C is the second-order cross feature vector, where c = [c (1,2) ,...,c (k-1,k) ].

[0095] Here, the second-order cross-feature vector is used as the input to the click behavior prediction network. The second-order feature vector is calculated by the inner product of the first-order feature vectors. Compared with the traditional method of using both the first-order and second-order feature vectors as input to the click behavior prediction network, the second-order feature vector is a scalar and has a smaller dimension, which can reduce network parameters and computational complexity. Furthermore, the click behavior prediction network can focus on extracting higher-order features, while lower-order features can be extracted in subsequent steps to improve the overall performance of the click behavior prediction model.

[0096] In step S13, the first-order feature vector is input into the feature weight layer of the click behavior prediction model to obtain a target weight matrix corresponding to the first-order feature vector, and a weighted first-order feature scalar is output according to the target weight matrix.

[0097] The click behavior prediction model in the embodiment of the present invention includes not only a click behavior prediction network but also a feature weighted layer. The feature weighted layer includes a weight matrix, which includes multiple weight parameters. Each weight parameter can be adjusted based on the data characteristics of the input feature weighted layer to adapt to different data distributions.

[0098] The present invention inputs the first-order feature vector into the feature weight layer, and the feature weight layer matches a suitable weight parameter to each first-order feature in the first-order feature vector, thereby obtaining a target weight matrix corresponding to the first-order feature vector.

[0099] Traditional click behavior prediction models use both first-order and second-order feature vectors as input to the click behavior prediction network. The first-order feature vectors will continue to cross in the network to explore the connections between learning features, but this will weaken the memory ability of features that have a strong impact on the prediction results, that is, the memory ability of the current data distribution will be weakened.

[0100] The present invention inputs the first-order feature vector into the feature weight layer, thereby supplementing the first-order feature data that is not input into the click behavior prediction network. The first-order feature data does not undergo complex crosstalk. The target weight vector obtained from the first-order feature data enables the model to remember some very important features and better adapt to the current data distribution.

[0101] In addition, in order to reduce the computational complexity of the model, the first-order feature vector can be converted into a first-order feature scalar in the feature weight layer, thereby outputting a weighted first-order feature scalar.

[0102] In this embodiment of the present invention, each first-order feature vector is assigned a weight parameter, which can be dynamically adjusted based on the data distribution characteristics of online accounts and target multimedia resources. Specifically, when new sample data is collected during offline or online training, the weights corresponding to the first-order features can be adjusted and updated.

[0103] In step S14, the initial prediction vector and the weighted first-order feature scalar are concatenated and input into the output layer of the click behavior prediction model to obtain the click behavior prediction result of the account on the target multimedia resource.

[0104] The initial prediction vector and all first-order feature scalars are spliced to the output layer of the click behavior prediction model to obtain the predicted click rate of the account for the target multimedia resource. This predicted click rate is the click behavior prediction result. As shown in formula (2):

[0105]

[0106] Among them, y CTR is the predicted click rate of the account-target multimedia resource output by the model, is the learnable parameter of the click behavior prediction model, x dnn is the output of the click behavior prediction network, is the weighted first-order characteristic scalar.

[0107] Specifically, the initial prediction vector and the weighted first-order feature scalar are spliced together as the input of the click behavior prediction model. After passing through the activation function in the click behavior prediction model, a prediction value between 0 and 1 can be obtained. This prediction value is the click behavior prediction result.

[0108] Figure 2 is a schematic diagram of a click behavior prediction method provided according to an exemplary embodiment.

[0109] Reference Figure 2 , the left side shows the operation process of the click behavior prediction network, and the right side shows the operation process of the feature weight layer.

[0110] Specifically, in the click behavior prediction network, the second-order cross feature vector c (1,2) ,...,c (k-1,k) Input the click behavior prediction network DNN (Deep Neural Network) to obtain the initial prediction vector x dnnIn the right feature weight layer, the first-order feature vector [e1, e2..., e k-1 , e k ] Input the feature weight layer, the feature weight layer adjusts the weight parameters according to the data distribution of the first-order features, and obtains the target weight parameters [α1, α2..., α k-1 , α k ]. Set the target weight parameters [α1, α2..., α k-1 , α k ] is multiplied by the first-order characteristic scalar corresponding to the first-order eigenvector to obtain the weighted first-order characteristic scalar

[0111] The initial prediction vector x dnn and the weighted first-order eigenvalues Splice and input the output layer of the click behavior prediction network to obtain the final click behavior prediction result y CTR .

[0112] In summary, the embodiment of the present invention inputs the second-order cross feature vector into the click behavior prediction network, which reduces the input data dimension and the amount of calculation. In addition, this solution adds a feature weight layer and inputs the first-order feature vector into the feature weight layer. On the one hand, the first-order feature vector does not undergo complex crossover and can directly remember the characteristics of the data, thereby improving the model's ability to remember feature data. On the other hand, the feature weight layer can adjust the weight parameters according to the changes in the first-order feature vector, so that the model can adapt well to different data distributions. The increased computational complexity and spatial complexity of the model are both linear. Therefore, this solution improves the model's ability to remember feature data without increasing the computational complexity of the model too much, thereby further improving the accuracy of the prediction.

[0113] In a possible implementation, based on the above embodiment, as Figure 3 As shown, the above step S11 may include steps S111 to S112:

[0114] S111. Acquire characteristic information of an account and characteristic information of a target multimedia resource, where the target multimedia resource is a multimedia resource to be recommended to the account;

[0115] When the server detects an account refreshing an app page, searching for multimedia resources, or opening a page related to an app, it can be considered a request for the target multimedia resource. At this point, the server retrieves the account's profile information and the profile information of the target multimedia resource to be recommended to the account.

[0116] The characteristic information of an account may include the account ID, age, gender, city, favorite content tags, disliked content tags, average viewing time, device ID, etc. The characteristic information of a target multimedia resource may include the target multimedia resource ID, resource category, content tags, duration, etc.

[0117] The characteristic information of the account and the target multimedia resource can be selected and preset according to the needs.

[0118] S112 , calling a click behavior prediction model to construct a first-order feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource, and to construct a second-order cross feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource.

[0119] The click behavior prediction model includes layers for calculating first-order feature vectors and second-order cross-feature vectors.

[0120] Specifically, using the click behavior prediction model, the account's feature information and the target multimedia resource's feature information are encoded and mapped into vectors, resulting in first-order feature vectors corresponding to the account's feature information and the target multimedia resource's feature information. All feature information for an account corresponds to one first-order feature vector, and all feature information for a target multimedia resource also corresponds to one first-order feature vector.

[0121] Furthermore, every two first-order eigenvectors are combined to obtain a plurality of second-order cross eigenvectors. The combining method may include multiplication or inner product.

[0122] In step S111-step S112, the first-order feature vectors corresponding to the feature information of the account and the feature information of the target multimedia resource are constructed, and the original feature information of the account and the target multimedia resource can be obtained; and since the first-order feature vectors cannot effectively represent some useful combination features that are difficult to discover manually or obtain through simple data analysis, the second-order cross feature vectors corresponding to every two first-order feature vectors are constructed. The first-order feature vectors can be combined through the second-order cross feature vectors to explore the correlation between the features.

[0123] In a possible implementation, based on the above embodiment, as Figure 4 As shown, the above step S112 may include steps S1121 to S1125:

[0124] S1121. Divide the feature information of the account and the feature information of the target multimedia resource into discrete features and continuous features.

[0125] Discrete features are features whose values can only be described by natural numbers, integers, counting units, etc. For example, account ID, age, gender, city, etc.

[0126] Continuous features are features that can take any value within a certain range, and two adjacent values can be divided infinitely. Examples include content tags that an account likes and dislikes, and content tags of target multimedia resources.

[0127] S1122. Input the discrete features into the embedding layer to obtain a first dense vector.

[0128] The feature information of the account and the feature information of the target multimedia resource are encoded using one-hot encoding. This data format is very sparse. If the sparse features are directly input into the neural network, it will cause dimensionality explosion. Therefore, the sparse features can be converted into dense features through word embedding.

[0129] Specifically, the click behavior prediction model also includes an embedding layer. The discrete features of the account and the target multimedia resource are input into the embedding layer to obtain a low-dimensional first dense vector.

[0130] If there are (k-1) discrete features of the account and target multimedia resource input to the embedding layer, then (k-1) first dense vectors are obtained, which can be expressed as: [e1, e2..., e k-1 ].

[0131] S1123: Concatenate the continuous features of the account and the continuous features of the target multimedia resource and input them into the perception layer to obtain a second dense vector with the same dimension as the first dense vector.

[0132] After passing through the embedding layer, discrete features will become high-dimensional vectors such as 16-dimensional and 32-dimensional, while continuous features are a single value and cannot maintain the same dimension as discrete features, and vectors of different dimensions cannot be cross-combined.

[0133] The click behavior prediction model also includes a perception layer. To facilitate subsequent cross-combination, N consecutive features can be concatenated to obtain an N-dimensional vector. This N-dimensional vector is then input into the perception layer to obtain a higher-dimensional vector of, for example, 16 or 32 dimensions.

[0134] In this way, the vector dimensions of continuous features and discrete features become consistent, which facilitates subsequent processing.

[0135] Specifically, for the continuous features of the account and the target multimedia resource, all the continuous features are spliced together and input into the perception layer (MLP Layer) to obtain the same [e1, e2..., e k-1 ] A second dense vector e of the same dimension k .

[0136] S1124: Concatenate the first dense vector and the second dense vector to obtain a first-order feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource.

[0137] The first dense vector representing the discrete features and the second dense vector representing the continuous features are concatenated to obtain the first-order feature vector corresponding to the account feature information and the target multimedia resource feature information. The first-order feature vector can be expressed as: [e1, e2..., e k-1 , e k ].

[0138] S1125. Input the first-order eigenvector into the inner product layer.

[0139] The click behavior prediction model also includes an inner product layer, which calculates the inner product between every two first-order feature vectors. Specifically, regardless of the features of the account or the target multimedia resource, all first-order feature vectors are fed into the inner product layer, and the inner product between every two first-order feature vectors is calculated.

[0140] S1126 . Calculate the inner product of every two first-order feature vectors using the inner product layer to obtain a second-order cross feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource.

[0141] The inner product layer uses the inner product method to combine every two first-order eigenvectors to obtain multiple second-order cross eigenvectors. The inner product between every two first-order eigenvectors is calculated. The specific calculation method is shown in formula (3):

[0142] c(i,j)= <e i ,e j > (3)

[0143] Among them, <,> represents inner product calculation, e i ,e j They represent the i-th first-order eigenvector and the j-th first-order eigenvector respectively, and c(i,j) represents the inner product of the i-th first-order eigenvector and the j-th first-order eigenvector.

[0144] For k first-order eigenvectors, after performing inner product operations on each of them, k(k-1) / 2 second-order cross eigenvectors are obtained.

[0145] The second-order cross eigenvector can be expressed as: c = [c (1,2) ,...,c (k-1,k) ].

[0146] In steps S1121 to S1126, the discrete features are input into the embedding layer to obtain a first dense vector, and the continuous features are spliced together and input into the perception layer to obtain a second dense vector with the same dimension as the first dense vector. In this way, the original sparse features are converted into dense features, which reduces the feature dimension and facilitates the subsequent input into the neural network for training. The dimension of the continuous features is also made the same as the dimension of the discrete features, which facilitates the subsequent crossover operation.

[0147] Figure 5 FIG. 1 is a flowchart showing a method for training a click behavior prediction model according to an exemplary embodiment. Figure 5 As shown, the click behavior prediction model training method includes the following steps.

[0148] In step S21, an initial model is obtained; the initial model includes an initial click behavior prediction network and an initial feature weight layer.

[0149] Unlike typical initial models that only include a click behavior prediction network, the initial model in this invention also includes an initial feature weight layer to specifically train feature weights. An existing click behavior prediction network can be selected as the initial click behavior prediction network. The model parameters in the initial model are all preset initial values.

[0150] In step S22 , a plurality of training samples are obtained, wherein the training samples include feature data of sample accounts and feature data of sample multimedia resources, as well as click data of the sample accounts on the sample multimedia resources.

[0151] Historical account and multimedia resource feature information collected during application execution, as well as click data on historical multimedia resources, can be used as training samples. To improve the accuracy and applicability of the model, the more recent the training samples, the better.

[0152] Optionally, the training samples can be divided into a training set, a validation set, and a test set. The training set is used to train and adjust model parameters, the validation set is used to verify model accuracy and adjust model hyperparameters, and the test set is used to verify the generalization ability of the model.

[0153] In step S23, the initial click behavior prediction network is trained using the second-order cross feature vector of the training sample and the click data to obtain the click behavior prediction network.

[0154] The training process involves obtaining the second-order cross-feature vectors of sample accounts and sample multimedia resources, using them as input to the click behavior prediction network. The network then performs initial calculations to output a predicted value. Based on the true values in the click data and the loss function, the network calculates the loss value for that parameter. The network then assigns new values to the parameters to minimize the loss value. Training concludes after repeated input, prediction, and parameter adjustment until the loss value falls below the network's preset loss threshold, resulting in a fully trained click behavior prediction network.

[0155] In step S24, the initial feature weight layer is trained using the first-order feature vector of the training sample, the click behavior prediction network, and the click data to obtain the feature weight layer.

[0156] The first-order feature vectors from the training samples are obtained and assigned an initial weight matrix to them, resulting in a weighted first-order feature scalar. This weighted first-order feature scalar is then combined with the prediction results from the click behavior prediction network and fed into the initial model to obtain a predicted value. Based on the difference between the predicted value and the true value in the click data, the weight parameters of the feature weight layer are adjusted, and the next prediction is performed until the loss value falls below the preset loss threshold for the feature weight layer. This concludes the training and results in a trained feature weight layer.

[0157] In step S25, a click behavior prediction model is obtained based on the click behavior prediction network and the feature weight layer.

[0158] After the click behavior prediction network and feature weight layer training are completed, the click behavior prediction model is obtained.

[0159] In summary, Figure 5 In the training method of the click behavior prediction model in , the initial model also includes an initial feature weight layer to specifically train the feature weights, which can improve the accuracy of weight parameter settings; and, determining the one-dimensional scalar corresponding to the first-order feature vector can reduce the dimension of the model and reduce the complexity of model operations; determining the normalized scalar corresponding to the one-dimensional scalar can avoid different dimensions and dimensions from offsetting the weight parameters, so that the weight can represent the importance of each feature, thereby further improving the accuracy of weight parameter settings.

[0160] In a possible implementation, the initial feature weight layer includes an initial weight matrix. Based on the above embodiment, Figure 6 As shown, the above step S24 may include steps S241 to S245:

[0161] S241: Obtain first-order feature vectors of the feature information of the sample account and the feature information of the sample multimedia resource.

[0162] S242. Determine a one-dimensional scalar corresponding to the first-order eigenvector, and determine a normalized scalar corresponding to the one-dimensional scalar.

[0163] In order to reduce the dimension of the model and reduce the complexity of model operations, the one-dimensional scalar corresponding to the first-order eigenvector can be calculated.

[0164] For example, in practice, if there are 400 first-order eigenvectors, each of which corresponds to 16 dimensions, concatenating all the first-order eigenvectors and connecting them to the click behavior prediction network would require concatenating 6,400-dimensional eigenvectors. This is very large, causing a huge increase in the model's computational complexity. However, if the scalar corresponding to each first-order eigenvector is calculated, then each first-order feature has only one dimension, and a 400-dimensional model can be constructed. Therefore, using one-dimensional scalars can reduce the model's computational complexity. The scalar corresponding to each first-order eigenvector can be calculated by summing all the eigenvectors.

[0165] Furthermore, a first-order eigenvector represents multiple features, each with different dimensions and scales. Subsequently, the one-dimensional scalar needs to be multiplied by a weight parameter. To prevent the weight parameter from being offset by the different dimensions and scales, and to ensure that the weights represent the importance of each feature, the one-dimensional scalars from the same batch of inputs can be normalized to obtain a normalized one-dimensional scalar.

[0166] S243: Taking the product of the normalized scalar and the preset initial weight matrix as the initial weighted first-order feature scalar.

[0167] The normalized scalar is multiplied by the preset initial weight matrix to obtain a weighted one-dimensional scalar. The calculation method of the weighted one-dimensional scalar can be expressed as follows using formula (4):

[0168]

[0169] in, is the i-th weighted one-dimensional scalar, a i It is e i The corresponding learnable initial weights, normal refers to normalization.

[0170] All the initial weighted one-dimensional scalars are concatenated to obtain the initial weighted first-order feature scalar, as shown in the following formula (5):

[0171]

[0172] in, is the initial weighted first-order characteristic scalar.

[0173] S244. Input the initial weighted first-order feature scalar and the output result of the click behavior prediction network into the output layer of the initial model to obtain an initial click behavior prediction result.

[0174] The initial weighted first-order feature scalar obtained in step B3 and the output result of the click behavior prediction network are spliced together and input into the output layer of the initial model to obtain the initial click behavior prediction result.

[0175] S245. Based on the click data and the initial click behavior prediction result, dynamically adjust the initial weight matrix to obtain the feature weight layer.

[0176] According to the difference between the initial click behavior prediction result and the actual value in the click data, the weight parameters of the feature weight layer are adjusted, and the next prediction is performed until the loss value is less than the loss threshold preset by the feature weight layer. The training ends and the trained feature weight layer is obtained.

[0177] In steps S241 to S245, determining the one-dimensional scalar corresponding to the first-order eigenvector can reduce the dimension of the model and reduce the complexity of the model operation; determining the normalized scalar corresponding to the one-dimensional scalar can avoid the different dimensions and dimensions from offsetting the weight parameters, so that the weight can represent the importance of each feature, thereby improving the accuracy of the weight parameter setting.

[0178] Optionally, step S242 includes the following steps S2421-S2422:

[0179] S2421. Perform a sum operation on the first-order eigenvector to obtain a one-dimensional scalar corresponding to the first-order eigenvector.

[0180] Specifically, a sum pooling operation may be performed on the first-order eigenvectors to obtain a one-dimensional scalar.

[0181] The first-order eigenvector includes multiple first-order features, and finding the sum of all first-order features is a summation operation.

[0182] S2422. Normalize the one-dimensional scalar to obtain a normalized scalar corresponding to the one-dimensional scalar; the normalized scalar is a scalar with a mean of 0 and a variance of 1.

[0183] Specifically, the normalization processing methods include minimum-maximum normalization, z-score, decimal scaling normalization, etc., which can be selected according to needs and are not specifically limited here.

[0184] In steps S2421 and S2422, the first-order eigenvectors are summed to obtain a one-dimensional scalar corresponding to the first-order eigenvectors. The one-dimensional scalar is then normalized to obtain a normalized scalar corresponding to the one-dimensional scalar. In the above method, using a one-dimensional scalar can reduce the computational complexity of the model; and calculating the normalized scalar of the one-dimensional scalar can prevent the weight parameters from being offset by different dimensions, thereby enabling the weights to accurately represent the importance of each feature.

[0185] Figure 7 The figure is a structural block diagram of a click behavior prediction device according to an exemplary embodiment.

[0186] like Figure 7 As shown, the click behavior prediction device 70 includes:

[0187] The feature vector acquisition module 71 is configured to acquire a first-order feature vector and a second-order cross feature vector corresponding to feature information of an account and feature information of a target multimedia resource to be recommended to the account;

[0188] An initial prediction module 72 is configured to execute inputting the second-order cross feature vector into a click behavior prediction network of a click behavior prediction model, and output an initial prediction vector of the click behavior of the account on the target multimedia resource;

[0189] a feature scalar determination module 73 configured to input the first-order feature vector into a feature weight layer of the click behavior prediction model, obtain a target weight matrix corresponding to the first-order feature vector, and output a weighted first-order feature scalar according to the target weight matrix;

[0190] The prediction result determination module 74 is configured to execute the concatenation of the initial prediction vector and the weighted first-order feature scalar and input the concatenated vector into the output layer of the click behavior prediction model to obtain the click behavior prediction result of the account on the target multimedia resource.

[0191] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0192] Figure 8 is a block diagram of an electronic device for click behavior prediction according to an exemplary embodiment, and its internal structure diagram can be as follows Figure 8As shown. The electronic device includes a processor, a memory, and a network interface connected via a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the electronic device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a method for predicting click behavior is implemented.

[0193] Those skilled in the art will understand that Figure 8 The structure shown in the figure is only a block diagram of a part of the structure related to the scheme of the present disclosure, and does not constitute a limitation on the electronic device to which the scheme of the present disclosure is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0194] In an exemplary embodiment, an electronic device is also provided, including: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the click behavior prediction method as in the embodiment of the present disclosure.

[0195] In an exemplary embodiment, a computer-readable storage medium is also provided. When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the click behavior prediction method of the embodiment of the present disclosure. The computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.

[0196] In an exemplary embodiment, a computer program product including instructions is further provided. When the computer program product is executed on a computer, the computer is caused to execute the click behavior prediction method in the embodiment of the present disclosure.

[0197] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, which can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0198] Other embodiments of the present disclosure will readily occur to those skilled in the art after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the present disclosure being indicated by the following claims.

[0199] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A click behavior prediction method, characterized in that: The method comprises: Obtaining a first-order feature vector and a second-order cross feature vector corresponding to feature information of an account and feature information of a target multimedia resource to be recommended to the account; Inputting the second-order cross feature vector into a click behavior prediction network of a click behavior prediction model, and outputting an initial prediction vector of the account's click behavior on the target multimedia resource; Inputting the first-order feature vector into the feature weight layer of the click behavior prediction model to obtain a target weight matrix corresponding to the first-order feature vector, and outputting a weighted first-order feature scalar according to the target weight matrix; splicing the initial prediction vector and the weighted first-order feature scalar and inputting the concatenated vector into the output layer of the click behavior prediction model to obtain a prediction result of the click behavior of the account on the target multimedia resource; The obtaining of the first-order feature vector and the second-order cross feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource includes: Acquiring characteristic information of an account and characteristic information of a target multimedia resource, wherein the target multimedia resource is a multimedia resource to be recommended to the account; Invoking a click behavior prediction model to construct a first-order feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource, and constructing a second-order cross feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource; The click behavior prediction model further includes an inner product layer, and constructing a second-order cross feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource includes: Inputting the first-order eigenvector into the inner product layer; The inner product layer is used to calculate the inner product of every two first-order feature vectors to obtain a second-order cross feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource.

2. The method according to claim 1, characterized in that The click behavior prediction model is trained by the following method: Obtaining an initial model; the initial model includes an initial click behavior prediction network and an initial feature weight layer; Acquire multiple training samples, where the training samples include feature data of sample accounts and feature data of sample multimedia resources, and click data of the sample accounts on the sample multimedia resources; Using the second-order cross feature vector of the training sample and the click data, the initial click behavior prediction network is trained to obtain the click behavior prediction network; Training the initial feature weight layer using the first-order feature vector of the training sample, the click behavior prediction network, and the click data to obtain the feature weight layer; A click behavior prediction model is obtained based on the click behavior prediction network and the feature weight layer.

3. The method according to claim 2, characterized in that The initial feature weight layer includes an initial weight matrix, and the training of the initial feature weight layer using the training sample and the click behavior prediction network includes: Obtaining first-order feature vectors of feature information of the sample account and feature information of the sample multimedia resource; Determine a one-dimensional scalar corresponding to the first-order eigenvector, and determine a normalized scalar corresponding to the one-dimensional scalar; The product of the normalized scalar and the preset initial weight matrix is used as the initial weighted first-order characteristic scalar; Inputting the initial weighted first-order feature scalar and the output result of the click behavior prediction network into the output layer of the initial model to obtain an initial click behavior prediction result; Based on the click data and the initial click behavior prediction result, the initial weight matrix is dynamically adjusted to obtain the feature weight layer.

4. The method according to claim 3, characterized in that The determining of the one-dimensional scalar corresponding to the first-order eigenvector and determining the normalized scalar corresponding to the one-dimensional scalar includes: Performing a sum operation on the first-order eigenvector to obtain a one-dimensional scalar corresponding to the first-order eigenvector; Normalization is performed on the one-dimensional scalar to obtain a normalized scalar corresponding to the one-dimensional scalar; the normalized scalar is a scalar with a mean of 0 and a variance of 1.

5. The method according to claim 1, wherein The click behavior prediction model further includes an embedding layer and a perception layer. The calling of the click behavior prediction model to construct a first-order feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource includes: dividing the feature information of the account and the feature information of the target multimedia resource into discrete features and continuous features; Inputting the discrete features into the embedding layer to obtain a first dense vector; splicing the continuous features of the account and the continuous features of the target multimedia resource and inputting them into the perception layer to obtain a second dense vector with the same dimension as the first dense vector; The first dense vector and the second dense vector are concatenated to obtain a first-order feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource.

6. A click behavior prediction device, characterized in that: The device comprises: a feature vector acquisition module configured to acquire a first-order feature vector and a second-order cross feature vector corresponding to feature information of an account and feature information of a target multimedia resource to be recommended to the account; an initial prediction module configured to execute inputting the second-order cross feature vector into a click behavior prediction network of a click behavior prediction model, and output an initial prediction vector of the click behavior of the account on the target multimedia resource; a feature scalar determination module configured to input the first-order feature vector into a feature weight layer of the click behavior prediction model, obtain a target weight matrix corresponding to the first-order feature vector, and output a weighted first-order feature scalar according to the target weight matrix; a prediction result determination module configured to execute concatenation of the initial prediction vector and the weighted first-order feature scalar and input the concatenated vector into an output layer of the click behavior prediction model to obtain a prediction result of the click behavior of the account on the target multimedia resource; The feature vector acquisition module is further configured to execute: Acquiring characteristic information of an account and characteristic information of a target multimedia resource, wherein the target multimedia resource is a multimedia resource to be recommended to the account; Invoking a click behavior prediction model to construct a first-order feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource, and constructing a second-order cross feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource; The click behavior prediction model further includes an inner product layer, and the feature vector acquisition module is further configured to perform: Inputting the first-order eigenvector into the inner product layer; The inner product layer is used to calculate the inner product of every two first-order feature vectors to obtain a second-order cross feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource.

7. The device according to claim 6, characterized in that The click behavior prediction model is trained by the following modules: An initial module acquisition module is configured to execute acquisition of an initial model; the initial model includes an initial click behavior prediction network and an initial feature weight layer; A training sample acquisition module is configured to acquire a plurality of training samples, wherein the training samples include feature data of a sample account and feature data of a sample multimedia resource, and click data of the sample account on the sample multimedia resource; A first training module is configured to train the initial click behavior prediction network using the second-order cross feature vector of the training sample and the click data to obtain the click behavior prediction network; A second training module is configured to train the initial feature weight layer using the first-order feature vector of the training sample, the click behavior prediction network, and the click data to obtain the feature weight layer; The model determination module is configured to execute the click behavior prediction network and the feature weight layer to obtain a click behavior prediction model.

8. The device according to claim 7, characterized in that The initial feature weight layer includes an initial weight matrix, and the second training module is further configured to perform: Obtaining first-order feature vectors of feature information of the sample account and feature information of the sample multimedia resource; Determine a one-dimensional scalar corresponding to the first-order eigenvector, and determine a normalized scalar corresponding to the one-dimensional scalar; The product of the normalized scalar and the preset initial weight matrix is used as the initial weighted first-order feature scalar; Inputting the initial weighted first-order feature scalar and the output result of the click behavior prediction network into the output layer of the initial model to obtain an initial click behavior prediction result; Based on the click data and the initial click behavior prediction result, the initial weight matrix is dynamically adjusted to obtain the feature weight layer.

9. The device according to claim 8, characterized in that The second training module is further configured to perform: Performing a sum operation on the first-order eigenvector to obtain a one-dimensional scalar corresponding to the first-order eigenvector; Normalization is performed on the one-dimensional scalar to obtain a normalized scalar corresponding to the one-dimensional scalar; the normalized scalar is a scalar with a mean of 0 and a variance of 1.

10. The device according to claim 6, characterized in that The click behavior prediction model further includes an embedding layer and a perception layer, and the feature vector acquisition module is further configured to perform: dividing the feature information of the account and the feature information of the target multimedia resource into discrete features and continuous features; Inputting the discrete features into the embedding layer to obtain a first dense vector; splicing the continuous features of the account and the continuous features of the target multimedia resource and inputting them into the perception layer to obtain a second dense vector with the same dimension as the first dense vector; The first dense vector and the second dense vector are concatenated to obtain a first-order feature vector corresponding to the feature information of the account and the feature information of the target multimedia resource.

11. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the click behavior prediction method according to any one of claims 1 to 5. 12 . A computer-readable storage medium, which, when instructions in the computer-readable storage medium are executed by a processor of a server, enables the server to execute the click behavior prediction method according to claim 1 .

13. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the click behavior prediction method according to any one of claims 1 to 5 is implemented.

Images